Bridging experiment and deep learning: Predicting of electronic properties in organic semiconductors using residual-gated graph neural networks

The adaptability of organic molecules offers an expansive design landscape for Organic semiconductors (OSCs) in organic solar cells (OSCells). OSCs are recognized for their outstanding electronic and optoelectronic properties critical for high-efficiency photovoltaic devices. Herein, we present a cutting-edge approach using Machine Learning (ML) and Deep Learning (DL) techniques to investigate energy gaps, Highest Occupied Molecular Orbital (HOMO), Lowest Unoccupied Molecular Orbital (LUMO), and electronic excitation energies, crucial for solar cells. The ML and DL models are rigorously trained on a dataset of over 48,182 OSCs derived from TD-DFT (Time-Dependent Density Functional Theory) calculations, utilizing SMILES as input features. Among these models, the residual-gated graph neural network (RGNN) emerges as the optimal model, achieving comparable accuracy to traditional DFT methods while dramatically reducing the computational costs, effectively predicting key properties. Our optimal DL model is validated against experimental data, results aligning closely with experimental findings, confirming the practical reliability of this approach for photovoltaic applications. Additionally, RGNN-based Web Server provides a computational platform with user-centric interface for rapid prediction of HOMO, LUMO, and excitation energy levels. This tool offers streamlined, accessible researchers, facilitating the synthesis of next-generation high-efficiency OSCs for solar cell applications, presenting the most reliable computational approach.